YOLO-BTM: A Novel Shuttlecock Detection Method for Embedded Badminton Robots

Abstract

Employing robots in badminton training contributes to a more accurate analysis of an athlete's movements and helps avoid injuries. Shuttlecock detection during the flying stage is a critical component of the badminton robot design. However, previous shuttlecock localization methods were unable to detect shuttlecock quickly and accurately in embedded device-based badminton robots, given scale variations, few extractable features, occlusion, and device limitation. In this paper, a deep learning-based shuttlecock localization method is proposed. First, an indoor shuttlecock dataset including 9548 shuttlecock images of various angles and scenes was constructed. Then a shuttlecock detection method YOLO-BTM is proposed, which is based on YOLOv4-Tiny. We proposed a new convolution block to replace the cross-stage partially block in the backbone, to improve the detection speed. To improve the network's ability to detect small objects, the efficient channel attention block is introduced in feature fusion. Finally, a comparative experiment on the accuracy of the method and the detection speed was conducted. The results show that the proposed YOLO-BTM has better performance in detection speed and accuracy compared to the existing state-of-the-art object detection methods on our own shuttlecock dataset. Our method enables real-time, accurate localization of shuttlecock and has the potential to be used in other embedded device based sports robots.